Background Remover

A threshold, a flood fill, and why that is enough

The tool first works out which color the background is: it looks at the pixels around the border, finds the shade most of them agree on, and uses that. You can override it with white, black, chroma-key green or any color you type in.

Then it walks inwards from the edges. Every pixel close enough to that color becomes transparent, and the walk stops at the first pixel that is not - which is what makes the flood fill smarter than a global color replacement. A white shirt on a white background survives, because the fill cannot reach it without crossing the subject. A green screen wants the other behavior, so there is a mode that removes the color anywhere in the picture instead.

The alpha channel is then multiplied by that mask, so the subject is not redrawn: the pixels that stay are exactly the pixels that were there. What you see in the preview is the file you download.

Tolerance is the one setting that matters

Photos are full of noise. A background that looks like one flat color is usually hundreds of near-identical shades, and the tolerance is how far a pixel may stray from the reference and still be treated as background. Too low and the background comes back as a speckled fringe; too high and the fill starts eating into the subject.

Distance is measured with the red-mean weighting from the color-difference literature rather than a plain RGB distance, because the eye is far more sensitive to green than to blue - which is what makes "20%" behave about the same way on a white background as on a green one.

Edge softness then blurs the mask, so the cut does not look like a staircase. Photographs usually want one or two pixels; a graphic or a scan with hard edges usually wants none. Because the mask is blurred and not the picture, the softened pixels stay in proportion to the subject instead of turning into a halo.

What this cannot do

It cannot separate hair from a background. Hair, fur, feathers and anything see-through are a blend of subject and background rather than a clean line, and no color threshold can invent the missing part - that is what segmentation models are for, and they cost tens of megabytes to download and still make mistakes.

It cannot remove a shadow, because a shadow is the same background color with the light taken out of it; lowering the tolerance enough to keep the subject usually leaves the shadow behind. It cannot tell "background" from "large area of the same color" - which is why the two spread modes exist, and why the tool tells you which one you used.

It also cannot put anything back: the removed pixels are gone, and the result is a PNG because a JPEG has nowhere to store transparency. Keep the original, and treat the output as the cut-out.

Trim, and the preview that shows the cut edge

The transparent border is a choice, not an accident. Trimming crops it away, so the PNG is exactly the size of the subject - right when the cut-out goes into a layout and has to fill its box. Keeping it leaves the picture the size of the original, which is what you want when several cut-outs have to line up in the same position.

The preview draws the result over a checkerboard, because only a checkerboard shows an edge honestly. You can spot a fringe or a missing piece before you download, and re-run with a new tolerance without ever closing a file.